Scholarlink

<p>Rapid detection of oil spills is essential to prevent severe damage to marine ecosystems, safeguard public health, and protect coastal economic activities. This study proposes ResNet-OSD, a hybrid deep learning framework designed for accurate oil spill detection using drone imagery. The architecture builds on ResNet50’s strong feature extraction capabilities to address key challenges such as complex shoreline textures, variable lighting, and environmental noise. Two hybrid configurations were developed and tested on drone images from oil-contaminated coastal areas. Both models applied techniques to improve learning and address class imbalance: SMOTE and Borderline SMOTE for oversampling, PCA for dimensionality reduction, and K-fold cross-validation for reliable evaluation. The first model, ResNet-SVM, combines ResNet50 with a support vector machine and uses SMOTE and K-fold validation. The second, ResNet-PCA-RF, integrates ResNet50 with principal component analysis and a random forest classifier, applying Borderline SMOTE for more targeted synthetic sampling. Experimental results show that ResNet-PCA-RF achieved superior performance, with an ROC–AUC of 1.00, accuracy of 98.43%, precision of 97.67%, and recall of 99.21%. These findings emphasize the model’s robustness and efficiency in classifying oil spills, even under challenging environmental conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ResNet-OSD: an optimized hybrid deep learning framework for oil spill detection in coastal drone imagery

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Iman A. Al-Sudani,
  • Ghaida A. Al-Suhail

摘要

Rapid detection of oil spills is essential to prevent severe damage to marine ecosystems, safeguard public health, and protect coastal economic activities. This study proposes ResNet-OSD, a hybrid deep learning framework designed for accurate oil spill detection using drone imagery. The architecture builds on ResNet50’s strong feature extraction capabilities to address key challenges such as complex shoreline textures, variable lighting, and environmental noise. Two hybrid configurations were developed and tested on drone images from oil-contaminated coastal areas. Both models applied techniques to improve learning and address class imbalance: SMOTE and Borderline SMOTE for oversampling, PCA for dimensionality reduction, and K-fold cross-validation for reliable evaluation. The first model, ResNet-SVM, combines ResNet50 with a support vector machine and uses SMOTE and K-fold validation. The second, ResNet-PCA-RF, integrates ResNet50 with principal component analysis and a random forest classifier, applying Borderline SMOTE for more targeted synthetic sampling. Experimental results show that ResNet-PCA-RF achieved superior performance, with an ROC–AUC of 1.00, accuracy of 98.43%, precision of 97.67%, and recall of 99.21%. These findings emphasize the model’s robustness and efficiency in classifying oil spills, even under challenging environmental conditions.